Abstract
Automatic emotion detection has become crucial in various domains, such as healthcare, neuroscience, smart home technologies, and human-computer interaction (HCI). Speech Emotion Recognition (SER) has attracted considerable attention because of its potential to improve conversational robotics and human-computer interaction (HCI) systems. Despite its promise, SER research faces challenges such as data scarcity, the subjective nature of emotions, and complex feature extraction methods. In this paper, we seek to investigate whether a lightweight deep neural ensemble model (CNN and CNN_Bi-LSTM) using well-known hand-crafted features such as ZCR, RMSE, Chroma STFT, and MFCC would outperform models that use automatic feature extraction techniques (e.g., spectrogram-based methods) on benchmarked datasets. The focus of this paper is on the effectiveness of careful fine-tuning of the neural models with learning rate (LR) schedulers and applying regularization techniques. Our proposed ensemble model is validated using five publicly available datasets: RAVDESS, TESS, SAVEE, CREMA-D, and EmoDB. Accuracy, AUC-ROC, AUC-PRC, and F1-score metrics were used for performance testing, and the LIME (Local Interpretable Model-agnostic Explanations) technique was used for interpreting the results of our proposed ensemble model. Results indicate that our ensemble model consistently outperforms individual models, as well as several compared models which include spectrogram-based models for the above datasets in terms of the evaluation metrics.